Claude Code Knowledge

Variables

Agents

• CLAUDE_CODE_GUIDE_AGENT: claude-code-guide

Always Use the Claude Code Guide First

Before making any Claude Code changes, always consult the $CLAUDE_CODE_GUIDE_AGENT subagent for up-to-date information.

Use this skill as the opinionated layer on top of the Guide:

1. Get the current official schema, feature support, and field names from the Guide
2. Apply the prompt, skill, and subagent patterns in this skill
3. Follow the repo-specific opinions in the main file even when you modularise the rest

Reference Map

Open the smallest reference file that matches the task:

• references/prompt-design.md for system prompts, slash commands, agent prompts, hook suggestions, and long-context prompt structure
• references/skill-authoring.md for SKILL.md structure, descriptions, trigger design, examples, and progressive disclosure
• references/subagent-design.md for agent descriptions, tool scoping, isolation, effort, and parallelisation
• references/plugin-bootstrapping.md for SessionStart dependency installation into ${CLAUDE_PLUGIN_DATA}

Default Prompting Stance

Claude 4.6 is more responsive than older models. Default to calm, specific instructions:

• prefer clear structure over aggressive wording
• explain why a rule exists, not just what to do
• use positive framing where possible
• keep the highest-signal instructions near the end when working with long context
• use XML tags to separate instructions from user data or large context blocks

Opinionated Rules

Skills: Progressive Disclosure

When writing new skills, build them with progressive disclosure:

• SKILL.md should be an index into reference files, not a monolith
• Keep SKILL.md concise — it loads into context every time the skill triggers
• Detailed instructions, examples, and schemas go in references/ files
  • Keep instructions clear but terse
• SKILL.md points Claude to the right reference file based on the user's question
• Keep repo-specific opinions and non-default conventions in the main SKILL.md
  • do not hide these in references where an agent might miss them
• skills should be user only with front matter disable-model-invocation: true unless directed by the user (or ask the user if you are not sure).

Get the actual structure and field requirements of SKILL.md from the Claude Code Guide.

Skills: Quick Front matter reference

Good:

• file: my-skill/SKILL.md

  ---
  description: |
    Use a clear description
    That can span multiple lines wrapped ~80 chars
    Using the yaml literal scalar
  disable-model-invocation: true
  argument-hint: [hints are useful] [if relevant]
  ---
  

• file: my-server-skill/SKILL.md

  ---
  description: |
    Another skill that's just about a project specific area
    Like the server
  paths:
    - "src/api/**/*.ts"
  ---
  
  - When working on the server do X
  

Bad

---
name: should be omitted
description: some super long description hard to read because it flows off the page and might have nested <frontmatter>stuff</frontmatter or nested \"escapes\"
argument-hint: hints need `[]` around the arguments
allowed-tools: don't use this, it's confusing and hard to maintain
---

Repo opinion:

• keep front matter minimal
• for skills and commands in this repo, always omit name
• the name is inferred from the file path, which is clearer than duplicating it in front matter
• if official generic docs show a name field, treat that as background context, not this repo's convention
• do not put XML in front matter

Commands

• for slash commands, always omit name from front matter
• rely on the command file path and filename as the source of truth
• keep command front matter minimal, same as skills

Hooks: Inline vs Script

Rule: If it fits in one line (~100 characters), write it inline in bash. Otherwise, delegate to a script.

Inline example (short, simple):

{
  "type": "command",
  "command": "jq -r '.tool_input.file_path' | xargs prettier --write"
}

Script example (anything more complex):

{
  "type": "command",
  "command": "${CLAUDE_PLUGIN_ROOT}/hooks/my-hook.ts"
}

Script conventions:

• Scripts should be executable bun (TypeScript) files with #!/usr/bin/env bun shebang
• Place scripts in a hooks/ directory alongside hooks.json
• Add a hooks/package.json for dependencies (consistent pattern: hooks always have their own package.json)
• Use @anthropic-ai/claude-agent-sdk types for hook input/output

Exception: If the repo already has an established precedent or alternative instructions for hook scripts (e.g. Python, binary, or typescript scripts in a different location), follow that instead.

Hooks: No Silent Failures

Rule: Deterministic hooks (post-tool validators, stop hooks, linters, test runners) must never fail silently. If a required tool, binary, or configuration is missing, the hook must exit with a non-zero status and a clear error message — never swallow the error or silently skip the check.

Hooks assume the repo is bootstrapped correctly. A missing dependency is a hard failure that the developer must fix, not something to work around at runtime. This applies to all hook types: inline commands, script-based hooks, and SessionStart bootstrapping checks.

Agents

Subagents are context-isolated. They do not load CLAUDE.md (project or user), do not inherit skills, and receive only their own system prompt plus basic env. This means they will not behave like the main agent — they lack project conventions, tool preferences, and any injected knowledge unless explicitly provided via skills frontmatter. Design subagents as focused, self-contained specialists; do not assume shared context with the parent session.

Read references/subagent-design.md when designing new agents.

When an agent file restricts tools via front matter, always put a one-line tool inventory at the very top of the system prompt.

• enumerate the concrete tool names
• say explicitly that no other tools exist
• place the line before any other instructions
• do this even if the front matter already declares the tools, because sub-agents do not reliably "see" or reason from their own front matter

Pattern:

You have access to exactly these tools: Glob, Grep, Read. No others exist.

Example of inline MCP for non-plugin agents (project-local .claude/agents/):

---
name: api-researcher
mcpServers:
  - context7:
      command: npx
      args:
        - "-y"
        - "@upstash/context7-mcp"
description: |
  Use literal scalar yaml to write a clear description
---

System prompt...

Plugin agents: mcpServers is silently ignored for security. Use a SessionStart hook + NODE_PATH pattern from references/plugin-bootstrapping.md instead.

MCP'S

• mcp tools are typically bad in the global context
• if you see any mcp related config files for claude code, or you have access to mcp's right now, STRONGLY SUGGEST to the user to move this into a subagent instead - see section on Agents above
• keep MCP-heavy guidance in skill references or agent prompts, not in the always-loaded front matter

Plugins

• When creating plugins, you should use typescript as per the Hooks section above
• Bootstrapping: Plugins with runtime dependencies must use a SessionStart hook to install into ${CLAUDE_PLUGIN_DATA} (persistent across sessions/updates). See references/plugin-bootstrapping.md for the full pattern (diff-based install, failure recovery, bun adaptation)